MultiMatte

What it does

MultiMatte is an image background removal model that uses text prompts to determine what stays and what gets removed. A user names an object in an image—"the dog" or "the jeans"—and the model keeps only that object while removing the rest. It handles edges that SAM 3 struggles with, like hair and translucent surfaces, by assigning opacity values to pixels rather than binary keep-or-remove decisions.

Who it is for

The tool targets developers and companies that need programmatic image segmentation. It works as an open source model, so engineering teams building image processing pipelines, design tools, or content platforms could integrate it. The model is available for testing at usefeyn.com/multimatte on individual images.

Pricing

The site does not show prices.

How it stands out

MultiMatte improves on Meta's SAM 3 by fine-tuning only 2.27% of the model's 860M parameters using low-rank adaptation. On the DIS-VD benchmark, it scores 0.901 S-measure versus SAM 3's 0.667. The key difference is the shift from binary segmentation masks to alpha mattes: instead of marking pixels as simply inside or outside an object, alpha mattes assign continuous opacity values. This lets the model preserve fuzzy or semi-transparent boundaries that binary masks cannot capture. Training used 19,953 images across diverse scenarios—salient objects, camouflage, high-resolution subjects, hair, and marine scenes—with 24.8% labeled with human-written object descriptions.

What a founder should check

A competitor would need to verify how much existing demand exists for API-based background removal versus the free or low-cost tools already available from Adobe, Photoshop, and other mainstream software. The second consideration is switching cost: developers who have already integrated SAM 3 or other segmentation models into production may not migrate unless MultiMatte's performance gains materially reduce errors in their use case. Third, assess whether the 2.27% fine-tuning approach creates meaningful defensibility. If competitors can replicate the LoRA technique on SAM 3 or apply it to newer foundation models, the moat shrinks quickly. The open source release also means anyone can download and modify the weights.